1use serde::{Deserialize, Serialize};
2use strum::EnumIter;
3
4#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
5#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
6pub enum BedrockModelMode {
7 #[default]
8 Default,
9 Thinking {
10 budget_tokens: Option<u64>,
11 },
12}
13
14#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
15#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
16pub struct BedrockModelCacheConfiguration {
17 pub max_cache_anchors: usize,
18 pub min_total_token: u64,
19}
20
21#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
22#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, EnumIter)]
23pub enum Model {
24 // Anthropic models (already included)
25 #[default]
26 #[serde(rename = "claude-sonnet-4", alias = "claude-sonnet-4-latest")]
27 ClaudeSonnet4,
28 #[serde(
29 rename = "claude-sonnet-4-thinking",
30 alias = "claude-sonnet-4-thinking-latest"
31 )]
32 ClaudeSonnet4Thinking,
33 #[serde(rename = "claude-opus-4", alias = "claude-opus-4-latest")]
34 ClaudeOpus4,
35 #[serde(
36 rename = "claude-opus-4-thinking",
37 alias = "claude-opus-4-thinking-latest"
38 )]
39 ClaudeOpus4Thinking,
40 #[serde(rename = "claude-3-5-sonnet-v2", alias = "claude-3-5-sonnet-latest")]
41 Claude3_5SonnetV2,
42 #[serde(rename = "claude-3-7-sonnet", alias = "claude-3-7-sonnet-latest")]
43 Claude3_7Sonnet,
44 #[serde(
45 rename = "claude-3-7-sonnet-thinking",
46 alias = "claude-3-7-sonnet-thinking-latest"
47 )]
48 Claude3_7SonnetThinking,
49 #[serde(rename = "claude-3-opus", alias = "claude-3-opus-latest")]
50 Claude3Opus,
51 #[serde(rename = "claude-3-sonnet", alias = "claude-3-sonnet-latest")]
52 Claude3Sonnet,
53 #[serde(rename = "claude-3-5-haiku", alias = "claude-3-5-haiku-latest")]
54 Claude3_5Haiku,
55 Claude3_5Sonnet,
56 Claude3Haiku,
57 // Amazon Nova Models
58 AmazonNovaLite,
59 AmazonNovaMicro,
60 AmazonNovaPro,
61 AmazonNovaPremier,
62 // AI21 models
63 AI21J2GrandeInstruct,
64 AI21J2JumboInstruct,
65 AI21J2Mid,
66 AI21J2MidV1,
67 AI21J2Ultra,
68 AI21J2UltraV1_8k,
69 AI21J2UltraV1,
70 AI21JambaInstructV1,
71 AI21Jamba15LargeV1,
72 AI21Jamba15MiniV1,
73 // Cohere models
74 CohereCommandTextV14_4k,
75 CohereCommandRV1,
76 CohereCommandRPlusV1,
77 CohereCommandLightTextV14_4k,
78 // DeepSeek
79 DeepSeekR1,
80 // Meta models
81 MetaLlama38BInstructV1,
82 MetaLlama370BInstructV1,
83 MetaLlama318BInstructV1_128k,
84 MetaLlama318BInstructV1,
85 MetaLlama3170BInstructV1_128k,
86 MetaLlama3170BInstructV1,
87 MetaLlama31405BInstructV1,
88 MetaLlama321BInstructV1,
89 MetaLlama323BInstructV1,
90 MetaLlama3211BInstructV1,
91 MetaLlama3290BInstructV1,
92 MetaLlama3370BInstructV1,
93 #[allow(non_camel_case_types)]
94 MetaLlama4Scout17BInstructV1,
95 #[allow(non_camel_case_types)]
96 MetaLlama4Maverick17BInstructV1,
97 // Mistral models
98 MistralMistral7BInstructV0,
99 MistralMixtral8x7BInstructV0,
100 MistralMistralLarge2402V1,
101 MistralMistralSmall2402V1,
102 MistralPixtralLarge2502V1,
103 // Writer models
104 PalmyraWriterX5,
105 PalmyraWriterX4,
106 #[serde(rename = "custom")]
107 Custom {
108 name: String,
109 max_tokens: u64,
110 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
111 display_name: Option<String>,
112 max_output_tokens: Option<u64>,
113 default_temperature: Option<f32>,
114 cache_configuration: Option<BedrockModelCacheConfiguration>,
115 },
116}
117
118impl Model {
119 pub fn default_fast(region: &str) -> Self {
120 if region.starts_with("us-") {
121 Self::Claude3_5Haiku
122 } else {
123 Self::Claude3Haiku
124 }
125 }
126
127 pub fn from_id(id: &str) -> anyhow::Result<Self> {
128 if id.starts_with("claude-3-5-sonnet-v2") {
129 Ok(Self::Claude3_5SonnetV2)
130 } else if id.starts_with("claude-3-opus") {
131 Ok(Self::Claude3Opus)
132 } else if id.starts_with("claude-3-sonnet") {
133 Ok(Self::Claude3Sonnet)
134 } else if id.starts_with("claude-3-5-haiku") {
135 Ok(Self::Claude3_5Haiku)
136 } else if id.starts_with("claude-3-7-sonnet") {
137 Ok(Self::Claude3_7Sonnet)
138 } else if id.starts_with("claude-3-7-sonnet-thinking") {
139 Ok(Self::Claude3_7SonnetThinking)
140 } else {
141 anyhow::bail!("invalid model id {id}");
142 }
143 }
144
145 pub fn id(&self) -> &str {
146 match self {
147 Model::ClaudeSonnet4 => "claude-4-sonnet",
148 Model::ClaudeSonnet4Thinking => "claude-4-sonnet-thinking",
149 Model::ClaudeOpus4 => "claude-4-opus",
150 Model::ClaudeOpus4Thinking => "claude-4-opus-thinking",
151 Model::Claude3_5SonnetV2 => "claude-3-5-sonnet-v2",
152 Model::Claude3_5Sonnet => "claude-3-5-sonnet",
153 Model::Claude3Opus => "claude-3-opus",
154 Model::Claude3Sonnet => "claude-3-sonnet",
155 Model::Claude3Haiku => "claude-3-haiku",
156 Model::Claude3_5Haiku => "claude-3-5-haiku",
157 Model::Claude3_7Sonnet => "claude-3-7-sonnet",
158 Model::Claude3_7SonnetThinking => "claude-3-7-sonnet-thinking",
159 Model::AmazonNovaLite => "amazon-nova-lite",
160 Model::AmazonNovaMicro => "amazon-nova-micro",
161 Model::AmazonNovaPro => "amazon-nova-pro",
162 Model::AmazonNovaPremier => "amazon-nova-premier",
163 Model::DeepSeekR1 => "deepseek-r1",
164 Model::AI21J2GrandeInstruct => "ai21-j2-grande-instruct",
165 Model::AI21J2JumboInstruct => "ai21-j2-jumbo-instruct",
166 Model::AI21J2Mid => "ai21-j2-mid",
167 Model::AI21J2MidV1 => "ai21-j2-mid-v1",
168 Model::AI21J2Ultra => "ai21-j2-ultra",
169 Model::AI21J2UltraV1_8k => "ai21-j2-ultra-v1-8k",
170 Model::AI21J2UltraV1 => "ai21-j2-ultra-v1",
171 Model::AI21JambaInstructV1 => "ai21-jamba-instruct-v1",
172 Model::AI21Jamba15LargeV1 => "ai21-jamba-1-5-large-v1",
173 Model::AI21Jamba15MiniV1 => "ai21-jamba-1-5-mini-v1",
174 Model::CohereCommandTextV14_4k => "cohere-command-text-v14-4k",
175 Model::CohereCommandRV1 => "cohere-command-r-v1",
176 Model::CohereCommandRPlusV1 => "cohere-command-r-plus-v1",
177 Model::CohereCommandLightTextV14_4k => "cohere-command-light-text-v14-4k",
178 Model::MetaLlama38BInstructV1 => "meta-llama3-8b-instruct-v1",
179 Model::MetaLlama370BInstructV1 => "meta-llama3-70b-instruct-v1",
180 Model::MetaLlama318BInstructV1_128k => "meta-llama3-1-8b-instruct-v1-128k",
181 Model::MetaLlama318BInstructV1 => "meta-llama3-1-8b-instruct-v1",
182 Model::MetaLlama3170BInstructV1_128k => "meta-llama3-1-70b-instruct-v1-128k",
183 Model::MetaLlama3170BInstructV1 => "meta-llama3-1-70b-instruct-v1",
184 Model::MetaLlama31405BInstructV1 => "meta-llama3-1-405b-instruct-v1",
185 Model::MetaLlama321BInstructV1 => "meta-llama3-2-1b-instruct-v1",
186 Model::MetaLlama323BInstructV1 => "meta-llama3-2-3b-instruct-v1",
187 Model::MetaLlama3211BInstructV1 => "meta-llama3-2-11b-instruct-v1",
188 Model::MetaLlama3290BInstructV1 => "meta-llama3-2-90b-instruct-v1",
189 Model::MetaLlama3370BInstructV1 => "meta-llama3-3-70b-instruct-v1",
190 Model::MetaLlama4Scout17BInstructV1 => "meta-llama4-scout-17b-instruct-v1",
191 Model::MetaLlama4Maverick17BInstructV1 => "meta-llama4-maverick-17b-instruct-v1",
192 Model::MistralMistral7BInstructV0 => "mistral-7b-instruct-v0",
193 Model::MistralMixtral8x7BInstructV0 => "mistral-mixtral-8x7b-instruct-v0",
194 Model::MistralMistralLarge2402V1 => "mistral-large-2402-v1",
195 Model::MistralMistralSmall2402V1 => "mistral-small-2402-v1",
196 Model::MistralPixtralLarge2502V1 => "mistral-pixtral-large-2502-v1",
197 Model::PalmyraWriterX4 => "palmyra-writer-x4",
198 Model::PalmyraWriterX5 => "palmyra-writer-x5",
199 Self::Custom { name, .. } => name,
200 }
201 }
202
203 pub fn request_id(&self) -> &str {
204 match self {
205 Model::ClaudeSonnet4 | Model::ClaudeSonnet4Thinking => {
206 "anthropic.claude-sonnet-4-20250514-v1:0"
207 }
208 Model::ClaudeOpus4 | Model::ClaudeOpus4Thinking => {
209 "anthropic.claude-opus-4-20250514-v1:0"
210 }
211 Model::Claude3_5SonnetV2 => "anthropic.claude-3-5-sonnet-20241022-v2:0",
212 Model::Claude3_5Sonnet => "anthropic.claude-3-5-sonnet-20240620-v1:0",
213 Model::Claude3Opus => "anthropic.claude-3-opus-20240229-v1:0",
214 Model::Claude3Sonnet => "anthropic.claude-3-sonnet-20240229-v1:0",
215 Model::Claude3Haiku => "anthropic.claude-3-haiku-20240307-v1:0",
216 Model::Claude3_5Haiku => "anthropic.claude-3-5-haiku-20241022-v1:0",
217 Model::Claude3_7Sonnet | Model::Claude3_7SonnetThinking => {
218 "anthropic.claude-3-7-sonnet-20250219-v1:0"
219 }
220 Model::AmazonNovaLite => "amazon.nova-lite-v1:0",
221 Model::AmazonNovaMicro => "amazon.nova-micro-v1:0",
222 Model::AmazonNovaPro => "amazon.nova-pro-v1:0",
223 Model::AmazonNovaPremier => "amazon.nova-premier-v1:0",
224 Model::DeepSeekR1 => "deepseek.r1-v1:0",
225 Model::AI21J2GrandeInstruct => "ai21.j2-grande-instruct",
226 Model::AI21J2JumboInstruct => "ai21.j2-jumbo-instruct",
227 Model::AI21J2Mid => "ai21.j2-mid",
228 Model::AI21J2MidV1 => "ai21.j2-mid-v1",
229 Model::AI21J2Ultra => "ai21.j2-ultra",
230 Model::AI21J2UltraV1_8k => "ai21.j2-ultra-v1:0:8k",
231 Model::AI21J2UltraV1 => "ai21.j2-ultra-v1",
232 Model::AI21JambaInstructV1 => "ai21.jamba-instruct-v1:0",
233 Model::AI21Jamba15LargeV1 => "ai21.jamba-1-5-large-v1:0",
234 Model::AI21Jamba15MiniV1 => "ai21.jamba-1-5-mini-v1:0",
235 Model::CohereCommandTextV14_4k => "cohere.command-text-v14:7:4k",
236 Model::CohereCommandRV1 => "cohere.command-r-v1:0",
237 Model::CohereCommandRPlusV1 => "cohere.command-r-plus-v1:0",
238 Model::CohereCommandLightTextV14_4k => "cohere.command-light-text-v14:7:4k",
239 Model::MetaLlama38BInstructV1 => "meta.llama3-8b-instruct-v1:0",
240 Model::MetaLlama370BInstructV1 => "meta.llama3-70b-instruct-v1:0",
241 Model::MetaLlama318BInstructV1_128k => "meta.llama3-1-8b-instruct-v1:0",
242 Model::MetaLlama318BInstructV1 => "meta.llama3-1-8b-instruct-v1:0",
243 Model::MetaLlama3170BInstructV1_128k => "meta.llama3-1-70b-instruct-v1:0",
244 Model::MetaLlama3170BInstructV1 => "meta.llama3-1-70b-instruct-v1:0",
245 Model::MetaLlama31405BInstructV1 => "meta.llama3-1-405b-instruct-v1:0",
246 Model::MetaLlama3211BInstructV1 => "meta.llama3-2-11b-instruct-v1:0",
247 Model::MetaLlama3290BInstructV1 => "meta.llama3-2-90b-instruct-v1:0",
248 Model::MetaLlama321BInstructV1 => "meta.llama3-2-1b-instruct-v1:0",
249 Model::MetaLlama323BInstructV1 => "meta.llama3-2-3b-instruct-v1:0",
250 Model::MetaLlama3370BInstructV1 => "meta.llama3-3-70b-instruct-v1:0",
251 Model::MetaLlama4Scout17BInstructV1 => "meta.llama4-scout-17b-instruct-v1:0",
252 Model::MetaLlama4Maverick17BInstructV1 => "meta.llama4-maverick-17b-instruct-v1:0",
253 Model::MistralMistral7BInstructV0 => "mistral.mistral-7b-instruct-v0:2",
254 Model::MistralMixtral8x7BInstructV0 => "mistral.mixtral-8x7b-instruct-v0:1",
255 Model::MistralMistralLarge2402V1 => "mistral.mistral-large-2402-v1:0",
256 Model::MistralMistralSmall2402V1 => "mistral.mistral-small-2402-v1:0",
257 Model::MistralPixtralLarge2502V1 => "mistral.pixtral-large-2502-v1:0",
258 Model::PalmyraWriterX4 => "writer.palmyra-x4-v1:0",
259 Model::PalmyraWriterX5 => "writer.palmyra-x5-v1:0",
260 Self::Custom { name, .. } => name,
261 }
262 }
263
264 pub fn display_name(&self) -> &str {
265 match self {
266 Self::ClaudeSonnet4 => "Claude Sonnet 4",
267 Self::ClaudeSonnet4Thinking => "Claude Sonnet 4 Thinking",
268 Self::ClaudeOpus4 => "Claude Opus 4",
269 Self::ClaudeOpus4Thinking => "Claude Opus 4 Thinking",
270 Self::Claude3_5SonnetV2 => "Claude 3.5 Sonnet v2",
271 Self::Claude3_5Sonnet => "Claude 3.5 Sonnet",
272 Self::Claude3Opus => "Claude 3 Opus",
273 Self::Claude3Sonnet => "Claude 3 Sonnet",
274 Self::Claude3Haiku => "Claude 3 Haiku",
275 Self::Claude3_5Haiku => "Claude 3.5 Haiku",
276 Self::Claude3_7Sonnet => "Claude 3.7 Sonnet",
277 Self::Claude3_7SonnetThinking => "Claude 3.7 Sonnet Thinking",
278 Self::AmazonNovaLite => "Amazon Nova Lite",
279 Self::AmazonNovaMicro => "Amazon Nova Micro",
280 Self::AmazonNovaPro => "Amazon Nova Pro",
281 Self::AmazonNovaPremier => "Amazon Nova Premier",
282 Self::DeepSeekR1 => "DeepSeek R1",
283 Self::AI21J2GrandeInstruct => "AI21 Jurassic2 Grande Instruct",
284 Self::AI21J2JumboInstruct => "AI21 Jurassic2 Jumbo Instruct",
285 Self::AI21J2Mid => "AI21 Jurassic2 Mid",
286 Self::AI21J2MidV1 => "AI21 Jurassic2 Mid V1",
287 Self::AI21J2Ultra => "AI21 Jurassic2 Ultra",
288 Self::AI21J2UltraV1_8k => "AI21 Jurassic2 Ultra V1 8K",
289 Self::AI21J2UltraV1 => "AI21 Jurassic2 Ultra V1",
290 Self::AI21JambaInstructV1 => "AI21 Jamba Instruct",
291 Self::AI21Jamba15LargeV1 => "AI21 Jamba 1.5 Large",
292 Self::AI21Jamba15MiniV1 => "AI21 Jamba 1.5 Mini",
293 Self::CohereCommandTextV14_4k => "Cohere Command Text V14 4K",
294 Self::CohereCommandRV1 => "Cohere Command R V1",
295 Self::CohereCommandRPlusV1 => "Cohere Command R Plus V1",
296 Self::CohereCommandLightTextV14_4k => "Cohere Command Light Text V14 4K",
297 Self::MetaLlama38BInstructV1 => "Meta Llama 3 8B Instruct",
298 Self::MetaLlama370BInstructV1 => "Meta Llama 3 70B Instruct",
299 Self::MetaLlama318BInstructV1_128k => "Meta Llama 3.1 8B Instruct 128K",
300 Self::MetaLlama318BInstructV1 => "Meta Llama 3.1 8B Instruct",
301 Self::MetaLlama3170BInstructV1_128k => "Meta Llama 3.1 70B Instruct 128K",
302 Self::MetaLlama3170BInstructV1 => "Meta Llama 3.1 70B Instruct",
303 Self::MetaLlama31405BInstructV1 => "Meta Llama 3.1 405B Instruct",
304 Self::MetaLlama3211BInstructV1 => "Meta Llama 3.2 11B Instruct",
305 Self::MetaLlama3290BInstructV1 => "Meta Llama 3.2 90B Instruct",
306 Self::MetaLlama321BInstructV1 => "Meta Llama 3.2 1B Instruct",
307 Self::MetaLlama323BInstructV1 => "Meta Llama 3.2 3B Instruct",
308 Self::MetaLlama3370BInstructV1 => "Meta Llama 3.3 70B Instruct",
309 Self::MetaLlama4Scout17BInstructV1 => "Meta Llama 4 Scout 17B Instruct",
310 Self::MetaLlama4Maverick17BInstructV1 => "Meta Llama 4 Maverick 17B Instruct",
311 Self::MistralMistral7BInstructV0 => "Mistral 7B Instruct V0",
312 Self::MistralMixtral8x7BInstructV0 => "Mistral Mixtral 8x7B Instruct V0",
313 Self::MistralMistralLarge2402V1 => "Mistral Large 2402 V1",
314 Self::MistralMistralSmall2402V1 => "Mistral Small 2402 V1",
315 Self::MistralPixtralLarge2502V1 => "Pixtral Large 25.02 V1",
316 Self::PalmyraWriterX5 => "Writer Palmyra X5",
317 Self::PalmyraWriterX4 => "Writer Palmyra X4",
318 Self::Custom {
319 display_name, name, ..
320 } => display_name.as_deref().unwrap_or(name),
321 }
322 }
323
324 pub fn max_token_count(&self) -> u64 {
325 match self {
326 Self::Claude3_5SonnetV2
327 | Self::Claude3Opus
328 | Self::Claude3Sonnet
329 | Self::Claude3_5Haiku
330 | Self::Claude3_7Sonnet
331 | Self::ClaudeSonnet4
332 | Self::ClaudeOpus4
333 | Self::ClaudeSonnet4Thinking
334 | Self::ClaudeOpus4Thinking => 200_000,
335 Self::AmazonNovaPremier => 1_000_000,
336 Self::PalmyraWriterX5 => 1_000_000,
337 Self::PalmyraWriterX4 => 128_000,
338 Self::Custom { max_tokens, .. } => *max_tokens,
339 _ => 128_000,
340 }
341 }
342
343 pub fn max_output_tokens(&self) -> u64 {
344 match self {
345 Self::Claude3Opus | Self::Claude3Sonnet | Self::Claude3_5Haiku => 4_096,
346 Self::Claude3_7Sonnet
347 | Self::Claude3_7SonnetThinking
348 | Self::ClaudeSonnet4
349 | Self::ClaudeSonnet4Thinking
350 | Self::ClaudeOpus4
351 | Model::ClaudeOpus4Thinking => 128_000,
352 Self::Claude3_5SonnetV2 | Self::PalmyraWriterX4 | Self::PalmyraWriterX5 => 8_192,
353 Self::Custom {
354 max_output_tokens, ..
355 } => max_output_tokens.unwrap_or(4_096),
356 _ => 4_096,
357 }
358 }
359
360 pub fn default_temperature(&self) -> f32 {
361 match self {
362 Self::Claude3_5SonnetV2
363 | Self::Claude3Opus
364 | Self::Claude3Sonnet
365 | Self::Claude3_5Haiku
366 | Self::Claude3_7Sonnet
367 | Self::ClaudeOpus4
368 | Self::ClaudeOpus4Thinking
369 | Self::ClaudeSonnet4
370 | Self::ClaudeSonnet4Thinking => 1.0,
371 Self::Custom {
372 default_temperature,
373 ..
374 } => default_temperature.unwrap_or(1.0),
375 _ => 1.0,
376 }
377 }
378
379 pub fn supports_tool_use(&self) -> bool {
380 match self {
381 // Anthropic Claude 3 models (all support tool use)
382 Self::Claude3Opus
383 | Self::Claude3Sonnet
384 | Self::Claude3_5Sonnet
385 | Self::Claude3_5SonnetV2
386 | Self::Claude3_7Sonnet
387 | Self::Claude3_7SonnetThinking
388 | Self::ClaudeOpus4
389 | Self::ClaudeOpus4Thinking
390 | Self::ClaudeSonnet4
391 | Self::ClaudeSonnet4Thinking
392 | Self::Claude3_5Haiku => true,
393
394 // Amazon Nova models (all support tool use)
395 Self::AmazonNovaPremier
396 | Self::AmazonNovaPro
397 | Self::AmazonNovaLite
398 | Self::AmazonNovaMicro => true,
399
400 // AI21 Jamba 1.5 models support tool use
401 Self::AI21Jamba15LargeV1 | Self::AI21Jamba15MiniV1 => true,
402
403 // Cohere Command R models support tool use
404 Self::CohereCommandRV1 | Self::CohereCommandRPlusV1 => true,
405
406 // All other models don't support tool use
407 // Including Meta Llama 3.2, AI21 Jurassic, and others
408 _ => false,
409 }
410 }
411
412 pub fn supports_caching(&self) -> bool {
413 match self {
414 // Only Claude models on Bedrock support caching
415 // Nova models support only text caching
416 // https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html#prompt-caching-models
417 Self::Claude3_5Haiku
418 | Self::Claude3_7Sonnet
419 | Self::Claude3_7SonnetThinking
420 | Self::ClaudeSonnet4
421 | Self::ClaudeSonnet4Thinking
422 | Self::ClaudeOpus4
423 | Self::ClaudeOpus4Thinking => true,
424
425 // Custom models - check if they have cache configuration
426 Self::Custom {
427 cache_configuration,
428 ..
429 } => cache_configuration.is_some(),
430
431 // All other models don't support caching
432 _ => false,
433 }
434 }
435
436 pub fn cache_configuration(&self) -> Option<BedrockModelCacheConfiguration> {
437 match self {
438 Self::Claude3_7Sonnet
439 | Self::Claude3_7SonnetThinking
440 | Self::ClaudeSonnet4
441 | Self::ClaudeSonnet4Thinking
442 | Self::ClaudeOpus4
443 | Self::ClaudeOpus4Thinking => Some(BedrockModelCacheConfiguration {
444 max_cache_anchors: 4,
445 min_total_token: 1024,
446 }),
447
448 Self::Claude3_5Haiku => Some(BedrockModelCacheConfiguration {
449 max_cache_anchors: 4,
450 min_total_token: 2048,
451 }),
452
453 Self::Custom {
454 cache_configuration,
455 ..
456 } => cache_configuration.clone(),
457
458 _ => None,
459 }
460 }
461
462 pub fn mode(&self) -> BedrockModelMode {
463 match self {
464 Model::Claude3_7SonnetThinking => BedrockModelMode::Thinking {
465 budget_tokens: Some(4096),
466 },
467 Model::ClaudeSonnet4Thinking => BedrockModelMode::Thinking {
468 budget_tokens: Some(4096),
469 },
470 Model::ClaudeOpus4Thinking => BedrockModelMode::Thinking {
471 budget_tokens: Some(4096),
472 },
473 _ => BedrockModelMode::Default,
474 }
475 }
476
477 pub fn cross_region_inference_id(&self, region: &str) -> anyhow::Result<String> {
478 let region_group = if region.starts_with("us-gov-") {
479 "us-gov"
480 } else if region.starts_with("us-") {
481 "us"
482 } else if region.starts_with("eu-") {
483 "eu"
484 } else if region.starts_with("ap-") || region == "me-central-1" || region == "me-south-1" {
485 "apac"
486 } else if region.starts_with("ca-") || region.starts_with("sa-") {
487 // Canada and South America regions - default to US profiles
488 "us"
489 } else {
490 anyhow::bail!("Unsupported Region {region}");
491 };
492
493 let model_id = self.request_id();
494
495 match (self, region_group) {
496 // Custom models can't have CRI IDs
497 (Model::Custom { .. }, _) => Ok(self.request_id().into()),
498
499 // Models with US Gov only
500 (Model::Claude3_5Sonnet, "us-gov") | (Model::Claude3Haiku, "us-gov") => {
501 Ok(format!("{}.{}", region_group, model_id))
502 }
503
504 // Available everywhere
505 (Model::AmazonNovaLite | Model::AmazonNovaMicro | Model::AmazonNovaPro, _) => {
506 Ok(format!("{}.{}", region_group, model_id))
507 }
508
509 // Models in US
510 (
511 Model::AmazonNovaPremier
512 | Model::Claude3_5Haiku
513 | Model::Claude3_5Sonnet
514 | Model::Claude3_5SonnetV2
515 | Model::Claude3_7Sonnet
516 | Model::Claude3_7SonnetThinking
517 | Model::ClaudeSonnet4
518 | Model::ClaudeSonnet4Thinking
519 | Model::ClaudeOpus4
520 | Model::ClaudeOpus4Thinking
521 | Model::Claude3Haiku
522 | Model::Claude3Opus
523 | Model::Claude3Sonnet
524 | Model::DeepSeekR1
525 | Model::MetaLlama31405BInstructV1
526 | Model::MetaLlama3170BInstructV1_128k
527 | Model::MetaLlama3170BInstructV1
528 | Model::MetaLlama318BInstructV1_128k
529 | Model::MetaLlama318BInstructV1
530 | Model::MetaLlama3211BInstructV1
531 | Model::MetaLlama321BInstructV1
532 | Model::MetaLlama323BInstructV1
533 | Model::MetaLlama3290BInstructV1
534 | Model::MetaLlama3370BInstructV1
535 | Model::MetaLlama4Maverick17BInstructV1
536 | Model::MetaLlama4Scout17BInstructV1
537 | Model::MistralPixtralLarge2502V1
538 | Model::PalmyraWriterX4
539 | Model::PalmyraWriterX5,
540 "us",
541 ) => Ok(format!("{}.{}", region_group, model_id)),
542
543 // Models available in EU
544 (
545 Model::Claude3_5Sonnet
546 | Model::Claude3_7Sonnet
547 | Model::Claude3_7SonnetThinking
548 | Model::ClaudeSonnet4
549 | Model::ClaudeSonnet4Thinking
550 | Model::Claude3Haiku
551 | Model::Claude3Sonnet
552 | Model::MetaLlama321BInstructV1
553 | Model::MetaLlama323BInstructV1
554 | Model::MistralPixtralLarge2502V1,
555 "eu",
556 ) => Ok(format!("{}.{}", region_group, model_id)),
557
558 // Models available in APAC
559 (
560 Model::Claude3_5Sonnet
561 | Model::Claude3_5SonnetV2
562 | Model::Claude3Haiku
563 | Model::Claude3Sonnet
564 | Model::Claude3_7Sonnet
565 | Model::Claude3_7SonnetThinking
566 | Model::ClaudeSonnet4
567 | Model::ClaudeSonnet4Thinking,
568 "apac",
569 ) => Ok(format!("{}.{}", region_group, model_id)),
570
571 // Any other combination is not supported
572 _ => Ok(self.request_id().into()),
573 }
574 }
575}
576
577#[cfg(test)]
578mod tests {
579 use super::*;
580
581 #[test]
582 fn test_us_region_inference_ids() -> anyhow::Result<()> {
583 // Test US regions
584 assert_eq!(
585 Model::Claude3_5SonnetV2.cross_region_inference_id("us-east-1")?,
586 "us.anthropic.claude-3-5-sonnet-20241022-v2:0"
587 );
588 assert_eq!(
589 Model::Claude3_5SonnetV2.cross_region_inference_id("us-west-2")?,
590 "us.anthropic.claude-3-5-sonnet-20241022-v2:0"
591 );
592 assert_eq!(
593 Model::AmazonNovaPro.cross_region_inference_id("us-east-2")?,
594 "us.amazon.nova-pro-v1:0"
595 );
596 Ok(())
597 }
598
599 #[test]
600 fn test_eu_region_inference_ids() -> anyhow::Result<()> {
601 // Test European regions
602 assert_eq!(
603 Model::ClaudeSonnet4.cross_region_inference_id("eu-west-1")?,
604 "eu.anthropic.claude-sonnet-4-20250514-v1:0"
605 );
606 assert_eq!(
607 Model::Claude3Sonnet.cross_region_inference_id("eu-west-1")?,
608 "eu.anthropic.claude-3-sonnet-20240229-v1:0"
609 );
610 assert_eq!(
611 Model::AmazonNovaMicro.cross_region_inference_id("eu-north-1")?,
612 "eu.amazon.nova-micro-v1:0"
613 );
614 Ok(())
615 }
616
617 #[test]
618 fn test_apac_region_inference_ids() -> anyhow::Result<()> {
619 // Test Asia-Pacific regions
620 assert_eq!(
621 Model::Claude3_5SonnetV2.cross_region_inference_id("ap-northeast-1")?,
622 "apac.anthropic.claude-3-5-sonnet-20241022-v2:0"
623 );
624 assert_eq!(
625 Model::Claude3_5SonnetV2.cross_region_inference_id("ap-southeast-2")?,
626 "apac.anthropic.claude-3-5-sonnet-20241022-v2:0"
627 );
628 assert_eq!(
629 Model::AmazonNovaLite.cross_region_inference_id("ap-south-1")?,
630 "apac.amazon.nova-lite-v1:0"
631 );
632 Ok(())
633 }
634
635 #[test]
636 fn test_gov_region_inference_ids() -> anyhow::Result<()> {
637 // Test Government regions
638 assert_eq!(
639 Model::Claude3_5Sonnet.cross_region_inference_id("us-gov-east-1")?,
640 "us-gov.anthropic.claude-3-5-sonnet-20240620-v1:0"
641 );
642 assert_eq!(
643 Model::Claude3Haiku.cross_region_inference_id("us-gov-west-1")?,
644 "us-gov.anthropic.claude-3-haiku-20240307-v1:0"
645 );
646 Ok(())
647 }
648
649 #[test]
650 fn test_meta_models_inference_ids() -> anyhow::Result<()> {
651 // Test Meta models
652 assert_eq!(
653 Model::MetaLlama370BInstructV1.cross_region_inference_id("us-east-1")?,
654 "meta.llama3-70b-instruct-v1:0"
655 );
656 assert_eq!(
657 Model::MetaLlama3170BInstructV1.cross_region_inference_id("us-east-1")?,
658 "us.meta.llama3-1-70b-instruct-v1:0"
659 );
660 assert_eq!(
661 Model::MetaLlama321BInstructV1.cross_region_inference_id("eu-west-1")?,
662 "eu.meta.llama3-2-1b-instruct-v1:0"
663 );
664 Ok(())
665 }
666
667 #[test]
668 fn test_mistral_models_inference_ids() -> anyhow::Result<()> {
669 // Mistral models don't follow the regional prefix pattern,
670 // so they should return their original IDs
671 assert_eq!(
672 Model::MistralMistralLarge2402V1.cross_region_inference_id("us-east-1")?,
673 "mistral.mistral-large-2402-v1:0"
674 );
675 assert_eq!(
676 Model::MistralMixtral8x7BInstructV0.cross_region_inference_id("eu-west-1")?,
677 "mistral.mixtral-8x7b-instruct-v0:1"
678 );
679 Ok(())
680 }
681
682 #[test]
683 fn test_ai21_models_inference_ids() -> anyhow::Result<()> {
684 // AI21 models don't follow the regional prefix pattern,
685 // so they should return their original IDs
686 assert_eq!(
687 Model::AI21J2UltraV1.cross_region_inference_id("us-east-1")?,
688 "ai21.j2-ultra-v1"
689 );
690 assert_eq!(
691 Model::AI21JambaInstructV1.cross_region_inference_id("eu-west-1")?,
692 "ai21.jamba-instruct-v1:0"
693 );
694 Ok(())
695 }
696
697 #[test]
698 fn test_cohere_models_inference_ids() -> anyhow::Result<()> {
699 // Cohere models don't follow the regional prefix pattern,
700 // so they should return their original IDs
701 assert_eq!(
702 Model::CohereCommandRV1.cross_region_inference_id("us-east-1")?,
703 "cohere.command-r-v1:0"
704 );
705 assert_eq!(
706 Model::CohereCommandTextV14_4k.cross_region_inference_id("ap-southeast-1")?,
707 "cohere.command-text-v14:7:4k"
708 );
709 Ok(())
710 }
711
712 #[test]
713 fn test_custom_model_inference_ids() -> anyhow::Result<()> {
714 // Test custom models
715 let custom_model = Model::Custom {
716 name: "custom.my-model-v1:0".to_string(),
717 max_tokens: 100000,
718 display_name: Some("My Custom Model".to_string()),
719 max_output_tokens: Some(8192),
720 default_temperature: Some(0.7),
721 cache_configuration: None,
722 };
723
724 // Custom model should return its name unchanged
725 assert_eq!(
726 custom_model.cross_region_inference_id("us-east-1")?,
727 "custom.my-model-v1:0"
728 );
729
730 Ok(())
731 }
732
733 #[test]
734 fn test_friendly_id_vs_request_id() {
735 // Test that id() returns friendly identifiers
736 assert_eq!(Model::Claude3_5SonnetV2.id(), "claude-3-5-sonnet-v2");
737 assert_eq!(Model::AmazonNovaLite.id(), "amazon-nova-lite");
738 assert_eq!(Model::DeepSeekR1.id(), "deepseek-r1");
739 assert_eq!(
740 Model::MetaLlama38BInstructV1.id(),
741 "meta-llama3-8b-instruct-v1"
742 );
743
744 // Test that request_id() returns actual backend model IDs
745 assert_eq!(
746 Model::Claude3_5SonnetV2.request_id(),
747 "anthropic.claude-3-5-sonnet-20241022-v2:0"
748 );
749 assert_eq!(Model::AmazonNovaLite.request_id(), "amazon.nova-lite-v1:0");
750 assert_eq!(Model::DeepSeekR1.request_id(), "deepseek.r1-v1:0");
751 assert_eq!(
752 Model::MetaLlama38BInstructV1.request_id(),
753 "meta.llama3-8b-instruct-v1:0"
754 );
755
756 // Test thinking models have different friendly IDs but same request IDs
757 assert_eq!(Model::ClaudeSonnet4.id(), "claude-4-sonnet");
758 assert_eq!(
759 Model::ClaudeSonnet4Thinking.id(),
760 "claude-4-sonnet-thinking"
761 );
762 assert_eq!(
763 Model::ClaudeSonnet4.request_id(),
764 Model::ClaudeSonnet4Thinking.request_id()
765 );
766 }
767}